Quickstart
This quickstart takes a new KGForEdGlobalMCP checkout from installation to a working local MCP connection.
The server runs over STDIO, loads the accepted curriculum graph packages from the repository, and exposes the fixed read-only MCP surface. The connected MCP host performs reasoning and generation; the server itself does not call an LLM.
What this quickstart verifies
By the end of this page, you will have synchronized the locked Python environment, exercised the real server through an MCP handshake, connected a local MCP host, and confirmed that the host can discover the curriculum catalog.
Prerequisites
Install the following before you begin:
| Requirement | Purpose |
|---|---|
| Git | Clone and update the repository |
uv |
Manage the required Python interpreter and locked environment |
| Python 3.13 | Runtime required by the backend (>=3.13,<3.14) |
| An MCP host | Connect to the local STDIO server; Claude Desktop is the default local integration |
Node.js, npm, and the MCPB CLI are needed only when building the optional .mcpb
distribution. They are not required for ordinary local server use.
Setup flow
%%{init: {"themeVariables": {"fontSize": "18px"}}}%%
flowchart LR
A[Clone repository] --> B[Install Python 3.13 with uv]
B --> C[Sync locked backend environment]
C --> D[Run STDIO smoke]
D --> E[Configure MCP host]
E --> F[List available frameworks]
1. Clone the repository
All commands on this page assume the repository root is your current directory.
2. Install the required Python version
The backend requires Python >=3.13,<3.14. Let uv manage the interpreter:
You do not need to activate a virtual environment manually when running the documented
commands through uv.
3. Synchronize the locked backend environment
Install the runtime dependencies from the backend lock file:
For ordinary server use, --no-dev keeps the environment limited to runtime
dependencies. Maintainers who also need development tooling can instead run:
See Local installation for the repository layout, environment variables, and development setup.
4. Verify the real STDIO server
Run the repository smoke command before configuring an MCP host:
The smoke command starts the real module entry point in a separate process, completes an MCP handshake, and checks the fixed public inventory:
- 13 tools;
- 1 fixed resource;
- 12 resource templates; and
- 7 prompts.
A successful run returns JSON containing:
Do not start the server by filesystem path
The supported local entry point is python -m kgfegmcp.mcpb_server. Directly
executing src/kgfegmcp/mcpb_server.py can cause the internal kgfegmcp.mcp package
to shadow the external MCP SDK package named mcp.
5. Connect an MCP host
The default local integration is Claude Desktop on macOS. It launches the same locked
backend through an absolute uv path and the module-based server entry point.
If you were given the URL of a hosted deployment instead, no local installation is needed; see Connect to a hosted server.
Follow Connect an MCP client for the complete configuration, JSON validation, restart, and connector-enablement steps.
6. Confirm framework discovery
After the connector is enabled in a new conversation, ask the host:
A successful response should identify accepted framework snapshots and preserve exact framework IDs, snapshot IDs, source metadata, and package capabilities.
Do not treat a host-generated summary as the source record itself. The MCP tools return the source-grounded evidence; the host may then summarize that evidence for you.
Success checklist
You are ready to use the server when all of the following are true:
uv python install 3.13succeeds;uv --directory backend sync --locked --no-devsucceeds;kgfegmcp-stdio-smokereports"status": "passed";- the MCP host shows the
curriculum-knowledge-graphconnector; and - a framework-discovery request returns the accepted catalog.
What the quickstart does not do
The quickstart does not:
- build curriculum graph packages from source documents;
- mutate accepted graph packages;
- create official alignments or learning progressions;
- enable semantic or embedding search; or
- run a server-side language model.
Those boundaries are intentional. See Concepts and boundaries before interpreting retrieved evidence across grades or frameworks.
Next: Local installation